Where Did I Come From? Origin Attribution of AI-Generated Images
Authors: Zhenting Wang, Chen Chen, Yi Zeng, Lingjuan Lyu, Shiqing Ma
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 5 Experiments and Results In this section, we first introduce the setup of the experiments ( 5.1). We evaluate the effectiveness of RONAN ( 5.2) and provide a case study on Stable Diffusion v2 model [3] ( 5.3). We then conduct ablation studies in 5.4, and discuss the comparison to existing reconstruction based attribution methods in 5.5. |
| Researcher Affiliation | Collaboration | Zhenting Wang Rutgers University zhenting.wang@rutgers.edu Chen Chen Sony AI Chen A.Chen@sony.com Yi Zeng Virginia Tech yizeng@vt.edu Lingjuan Lyu Sony AI Lingjuan.Lv@sony.com Shiqing Ma University of Massachusetts Amherst shiqingma@umass.edu |
| Pseudocode | Yes | Algorithm 1 Origin Attribution Input: Model: M, Examined Data: x Output: Inference Results: Belonging or Non-belonging 1: function INFERENCE(M, x) 2: Obtaining Belonging Distribution (Offline) 3: µ, σ, N = Belonging Distribution(M) 4: Reverse-engineering 5: A (M, x) Calibrated Reconstruction Loss [Eq. 2] 6: Determining Belonging 7: Inference Results = Hypothesis Testing(A (M, x), µ, σ, N)[Eq. 3] 8: return Inference Results |
| Open Source Code | Yes | Our code can be found in https://github.com/Zhenting Wang/RONAN. |
| Open Datasets | Yes | We conduct experiments on the CIFAR-10 [48] dataset using DCGAN [1], VAE [6], and Style GAN2-ADA [10] models. The investigated models include DCGAN [1], VAE [6], Style GAN2ADA [10] trained on the CIFAR-10 [48] dataset, Consistency Model [4] trained on the Image Net [53] dataset, and Control GAN [28] trained on the CUB-200-2011 [57] dataset. ... we randomly sample 10 images from MNIST [79] dataset... |
| Dataset Splits | No | The paper discusses distinguishing between belonging images and real images (training data or unseen data) and between belonging images and images generated by other models. It does not provide explicit details about a separate validation set split with percentages or counts. |
| Hardware Specification | Yes | We conducted all experiments on a Ubuntu 20.04 server equipped with six Quadro RTX 6000 GPUs. |
| Software Dependencies | Yes | Our method is implemented with Python 3.8 and Py Torch 1.11. |
| Experiment Setup | Yes | In line 3, we use the given model to generate N (i.e., 100 by default in this paper) images with randomly sampled inputs... α is the significance level of the hypothesis testing (i.e., 0.05 by default in this paper). In line 5, we calculate the calibrated reconstruction loss of the examined image (Eq. 2), the reconstruction loss is computed via gradient descent optimizer (Adam [55] by default in this paper). |